Professional Documents
Culture Documents
RESEARCH
R E P O R T
RUTCOR
Rutgers Center for
Operations Research
Rutgers University
640 Bartholomew Road
Piscataway, New Jersey
08854-8003
Telephone: 732-445-3804 a
hammer@rutcor.rutgers.edu , Tel.: +1 7324454812 ; Fax: +1 7324455472
b
Telefax: 732-445-5472 kogan@rutcor.rutgers.edu
c
Email: rrr@rutcor.rutgers.edu mlejeune@andromeda.rutgers.edu
a,b,c
http://rutcor.rutgers.edu/~rrr
RUTCOR - Rutgers University Center for Operations Research, Piscataway, NJ, USA
a,b
Rutgers Business School, Rutgers University, Newark-New Brunswick, NJ, USA
RUTCOR RESEARCH REPORT
RRR 8-2004, MARCH 2004
Abstract. The central objective of this paper is to develop transparent, consistent, self-
contained, and stable country risk rating systems, closely approximating the country risk ratings
provided by a major rating agency (Standard & Poor). We propose two models that achieve the
stated objectives, the first one utilizing the classical econometric technique of multiple linear
regression, and the second one using the combinatorial-logical technique of Logical Analysis of
Data. The proposed models use economic-financial and political variables, and are non-
recursive (i.e., they do not rely on the previous years’ ratings). The accuracy of the proposed
models’ predictions, measured by their correlation coefficients with Standard and Poor’s
ratings, and confirmed by k-folding cross-validation, exceeds 95%. The stability of the
constructed non-recursive models is shown in three ways: by the correlation of the predictions
with those of other agencies (Moody’s and The Institutional Investor), by predicting 1999
ratings using the non-recursive models derived from the 1998 dataset applied to the 1999 data,
and by successfully predicting the ratings of several previously non-rated countries. The
confidence in the results and in the validity of both models is strongly reinforced by the fact that
the traditional linear regression model and the qualitatively different combinatorial-logical
model produce almost identical results.
Acknowledgements: The authors express the appreciation to Dr. Sorin Alexe for his invaluable help in
the execution of computational experiments with LAD.
1 Country Risk, Country Risk Ratings and Objectives of the
Paper
default) the lower the interest rate. Following its sovereign rating downgrade , Japan’s borrowing
became more expensive as interest rates have increased, reflecting the higher chance of default ,
which deteriorates even more the situation of the heavily indebted Japanese government and
economy.
Second, sovereign ratings also influence credit ratings of national banks and companies, and
affect their attractiveness to foreign investors. Ferri et al. (2001) call sovereign ratings the “pivot
of all other country’s ratings”. Similarly, Erb et al. (1995a) underline that raters have historically
shown a reluctance to give a company a higher credit rating than that of the sovereign where the
company operates. For example, after Moody’s downgraded Japan in November 1998 (from Aaa
to Aa1), all other Aaa Japan issuers have been downgraded (Jüttner and McCarthy, 2000). This
led sovereign ratings to be named “sovereign credit risk ceilings”.
Third, institutional investors are sometimes contractually restricted on the degree of risk they
can assume, implying in particular that they cannot invest in debt rated below a prescribed level.
Ferri et al. (2001) refine this analysis, pointing out the contrast between the ratings of banks
operating in high- and low-income countries, and show that ratings of banks operating in low-
income countries are significantly affected by variations in sovereign ratings, while the ratings of
banks operating in high-income countries do not seem to depend significantly on country ratings.
Similarly, Kaminsky and Schmukler (2002) as well as Larrain et al. (1997) note that sovereign
ratings are crucial for developing economies, which have a very high sensitivity to rating
announcements.
Comprehensibility: The country risk ratings published by different agencies appear as outputs of
“black boxes”, the real content and meaning of which are unexplained and hard to understand,
since rating agencies specify neither the factors which are taken into consideration in determining
their ratings, nor the “rules of compression” of multiple factors into a single rating. This raised
the discontent of Japan’s Prime Minister, Junichiro Koizumi, who was “railed at being rated in
the same neighborhood as African countries to which Japan is providing assistance” . Officials of
Japan’s Ministry of Finance added that big rating agencies are “making unfair qualitative
judgments” , while Moody’s denied and claimed that the motives for the downgrade lie in the
“increased debt load” of Japan. In view of such controversy, uncovering both the factors which
are taken into account by these black boxes, and the mechanisms of deriving ratings, are essential
for ascertaining the consistency of a country rating system.
1
PAGE 2 RRR 08-2004
It is not clear however which ones of the many possible factors do actually influence the payback
capacity of a country. This question is subject to different analyses. Haque et al. (1998) claim that
it is sufficient to restrict the scope of analysis to economic/financial factors only, while others
(Brewer and Rivoli,1990) claim that both economic/financial and political factors impact country
risk ratings.
Rating failures: Some recent failures have challenged the trustworthiness of country risk ratings
(Reinhart, 2002, Levich et al., 2002) . Criticisms directed towards ratings institutions have been
especially intense after the Tequila and the Asian crises. Indeed, the tequila crisis in Mexico
(1994-95) had not been preceded by a rating downgrade, implying that either the crisis was not
predicted, or that its significance was overlooked. Similar observations apply to the Asian crisis
(1997-99): Ftich admitted that “it and its larger rivals Standard’s & Poor and Moody’s Investors
Services of the US had largely failed to predict the recent turmoil in Asia” . On the other hand,
rating agencies have been more insightful in anticipating other crises, e.g. in Russia (1998),
• Regional bias: Diverse explanations have been provided for the failure of rating
Brazil (1998) and Argentina (2001).
agencies to signal crisis emergencies in various countries. There are claims that certain
rating agencies favor certain regions. For instance, Haque et al. (1997) note that
Euromoney usually gives higher ratings to Asian and European countries than to Latin
or Caribbean countries, while the Institutional Investor is more generous to Asian and
• Latency: Another criticism lies in the time taken by the rating agencies to react to new
European countries than to African ones.
facts (e.g., according to The Economist , “rating agencies may have been too slow to
• Overreactions: The IMF criticizes rating agencies claiming that they reacted in panic
downgrade Japan. Markets have already moved ahead of them”).
during the Asian crisis. After they had missed to predict the Asian crisis, they reacted by
harshly downgrading countries such as Thailand or South Korea, thus accelerating the
flight of capital. In this and other situations, rating agencies gave the impression of
overreacting (Figure 1) instead of being a stabilizing force.
Figure 1: Precrisis and postcrisis rating of countries
The Institutional Investor's ratings
80 Precrisis 80 Precrisis
70 Postcrisis 70 Postcrisis
Euromoney's ratings
60 60
50 50
40 40
30 30
20 20
10 10
0 0
Thailand Korea Indonesia Thailand Korea Indonesia
It appears that the objectivity and reliability of country risk ratings is questionable, mainly
because of human intervention and conflicting goals and/or interests.
RRR 08-2004 PAGE 3
•
upgrade countries and intensify outflows of capital and crisis when they downgrade.
Conflicts of interest: An even more pointed criticism is that raters, having started
charging fees to rated countries, can be suspected of reluctance to downgrade them,
because of the possibility of jeopardizing their income sources. This is claimed, for
example, by Tom McGuire, an executive vice-president of Moody’s, who states that
“the pressure from fee-paying issuers for higher ratings must always be in a delicate
balance with the agencies’ need to retain credibility among investors”2. The necessity to
please the payers of the ratings, investors as well as issuers, lead to what Robert
Grossman, the chief credit officer at the rating agency Fitch, calls “a tendency we do
with investors – rating committees, outlooks, meetings, then the press release, all to
soften the blow of the rating change”3. Studying the rating transitions, Altman and
Saunders (1998) notice that a downgrade in the rating of a country is regularly followed
by further downward adjustments. The explanation given by Altman and Saunders is
that agencies gradually downgrade the rating of a country, since they do not want to hurt
the country, which is also their client. Kunczik (2001) note that the IMF (1999) fears the
danger that “issuers and intermediaries could be encouraged to engage in rating
shopping – a process in which the issuer searches for the least expensive and/or least
demanding rating”.
The problems described above will become more acute as the role of ratings increases. Indeed,
the Basel Accord will intensify the pressure on countries to obtain high ratings, potentially
leading to a switch from rating shopping to rating fraud. For instance, Pakistan has been forced to
pay back $55 million credits to the IMF because of budget falsification, the blame being put on
the former Prime Minister Nawaz Sharif, accused of having falsified the budget deficit. Similarly,
Ukraine has been proven to have reported misleading data on its reserves in foreign exchanges,
attempting to obtain IMF credits. Kunczik (2001) says that “it is only a question of time when
firms will specialize in rating advising for sovereigns”.
1
This article appeared in the February 19, 1999 issue Executive Intelligence Review.
Interview: Datuk Seri Dr. Mahathir bin Mohamad Malaysian Prime Minister: `We had to decide things for ourselves'.
On January 22, 1999, Gail G. Billington of EIR's Asia Desk and Dino de Paoli of the Schiller Institute were given the
opportunity to interview Datuk Seri Dr. Mahathir bin Mohamad, Prime Minister of Malaysia.
2
The Economist, July 15, 1995, 62
3
Euromoney, January 2002, 38, “Investors turn cool on the rating game”
3
PAGE 4 RRR 08-2004
500
400
300
200
100
0
June 1997 June 1998
It appears that the use of market spreads rather than country ratings is not more efficient.
Indeed, for Asian countries, spreads have substantially widened after the crisis. As exhibited by
Figure 2, spreads were roughly of the same order of magnitude before the crisis. While spreads of
non-crisis countries have widened by less than 100% after the crisis, spreads of crisis countries
have more than tripled. Consequently, we conclude that spreads provide about the same
information as sovereign ratings do, and are much more volatile. This conclusion can be extended
to the Brazilian and the Russian crises. This discussion implies that yield spreads are
characterized by a lack of predictive power and cannot be used to obtain a reliable early warning
of country insolvency. This latter conclusion is confirmed by Mathieson and Schinasi (1999).
4
T-bill rate, GDP growth rate, inflation rate, exports growth rate, ratio current account to GDP, the ratio of external
debt to GDP, the ratio of reserves to imports.
RRR 08-2004 PAGE 5
The 98% correlation level5 between The Institutional Investor ratings published respectively in
September 1997 and September 1998 confirms the stability property of sovereign ratings. In light
of this fact, the excellent correlation levels achieved by utilizing lagged ratings among the
independent variables can be attributed to a certain – possibly large – extent to this stability, and
may not necessarily give indications about the predictive power of the economic and political
variables used as predictors.
Although Cantor and Packer (1996) do not include the lagged ratings in their set of predictors,
they create a dummy variable, which is determined by the past ratings issued by Standard and
Poor (Claessens and Embrechts, 2002). This dummy variable is defined to be equal to 1 if a
country has ever been rated D or SD by Standard & Poor since 1970, and equal to 0 otherwise.
Even though their regression R-square is above 90%, their results are criticized by Claessens and
Embrechts (2002) and Jüttner and McCarthy (2000). Claessens and Embrechts mention that the
dates of the explanatory variables are not consistent, e.g. the values of some variables are
measured in 1994 or 1995, while that of others are averages for the period 1991-1994 or 1992-
1994. On the other hand, Jüttner and McCarthy evaluated the regression model of Cantor and
Packer for some other years, concluding that for 1998, it loses its predictive power. A recent
paper of Hu et al. (2002) develops a model using ordered probit to estimate country ratings. Their
model has an 83% correlation level and relies on economic variables and rating history of
countries.
A common feature of the econometric models above is the direct or indirect inclusion of
information derived from past Standard & Poor ratings (lagged ratings, rating history) among
their independent variables. A major drawback of such rating models is the impossibility of
applying them to not-yet-rated countries.
5
134 countries are considered.
5
PAGE 6 RRR 08-2004
In line with the existing literature, we use in the first part of this paper the technique of
multiple linear regression to achieve our objectives. We shall call the proposed system a non-
recursive multiple regression model of the S&P country rating system. In the second part of the
paper, we reanalyze the same problem, using this time a combinatorial-logical technique, in order
to derive a set of “logical rating scores” of countries, and show that they turn out to be
surprisingly similar to both the S&P ratings and the non-recursive regression scores.
The fact that the traditional linear regression model and the qualitatively different
combinatorial-logical model produce almost identical results strongly reinforces the confidence in
the results and in the validity of both models. In addition to the main result, we also demonstrate
political attributes of countries i and j which were identified in the first part of the study. An
additional component of a pseudo-observation is an indicator which takes the value 1 (-1, 0) if the
country i in the pseudo-observation has a higher (lower, identical) rating than the country j.
The fundamental idea of this study is that a rating system can be essentially reconstructed
from the knowledge of the relative orderings of all pairs of rated countries. In other words, all that
matters in a rating is the qualitative order relation between countries, but not a quantitative
measure of the magnitude of differences between ratings.
The study focuses on deriving a model of the order relation between countries using the LAD
methodology which is briefly described in Subsection 4.2. Non-statistical and highly nonlinear,
this methodology is not restricted by the satisfaction of the assumptions underlying econometric
Based on the patterns of the LAD model, a discriminant, ∆(Pij), called relative preference, is
techniques.
computed for each pseudo-observation, Pij, The value of ∆(Pij) indicates whether country i should
be rated higher or lower than country j. Relying on the assumption that the ∆(Pij) values provide
good approximations of the differences of the ratings, the relative preferences are used to derive
an approximation of the ratings called logical rating scores of countries; these are calculated
set of countries, the rating of a country k is the regression coefficient, β k , in the regression model
using multiple linear regression, as described in Subsection 4.3. More precisely, denoting by I the
2 Data
2.1 Sources
In this paper, we focus on the Standard & Poor country risk ratings. The risk of default is
generally defined by Standard & Poor as the probability that a sovereign obligor fails to meet a
principal or interest payment on the due date and in full. Standard & Poor’s ratings are based on
the information provided by the debtors themselves and by other sources considered reliable.
Standard & Poor provides sovereign ratings for local and foreign currency debt. In this paper,
we used the foreign currency sovereign ratings. Countries are more vulnerable to foreign
currency obligations. An obligor's capacity to repay foreign currency obligations may be lower
than its capacity to repay obligations in its local currency, owing to the sovereign government's
7
PAGE 8 RRR 08-2004
relatively lower capacity to repay external versus domestic debt. As noted by Cantor and Packer
(1996), foreign currency ratings remain the decisive factor in the international bond market.
Indeed, foreign currency obligations are more likely to be acquired by international investors than
domestic obligations. Foreign currency ratings reflect economic factors, as well as the country
intervention risk, i.e. the risk of a country imposing, for example, exchange controls or a debt
moratorium, while local currency ratings exclude country intervention risk.
Table 15 in the Appendix lists the different country risk levels or labels used by Standard &
Poor, and also provides descriptions associated with these labels. Countries which are assigned a
label inferior to BB+ are considered as non-investment grade (speculative) countries. Countries
rated CCC+ or lower are regarded as presenting serious default risks. BB indicates the least
degree of speculation and CC the highest. Ratings labeled from AA to CCC can be modified by
the addition of a plus or minus sign to show relative standing within the major rating categories.
We consider such subcategories as separate ratings in our analysis.
We have converted the Standard & Poor rating scale (ranging from AAA to SD) into a
numerical scale (ranging from 21 to 0) (see Appendix, Table 16) and shall liberally refer to both
of them as S&P ratings. This type of conversion is commonly used in the literature, see e.g.,
Bouchet et al. (2003), Estrella (2000), Ferri et al.(2001), Kräussl (2000), Monfort and Mulder
(2000), Mulder and Perelli (2001), Hu et al. (2002), Sy [2003]. Moreover, Bloomberg, a major
provider of financial data services, developed a standard cardinal scale for comparing Moody’s,
S&P and Fitch-BCA ratings (Kaminsky and Schmukler, 2002); in this scales, a higher numerical
value denotes a higher probability of default.
In Table 16 of the Appendix, we display Standard & Poor’s foreign currency sovereign ratings
of 69 countries published at the end of December 1998. Standard & Poor rates a limited number
of countries, with a special focus (at least in the past) on the industrial ones. However, in the last
decade, the number of Asian, Latin American and Eastern European economies rated by Standard
& Poor has significantly increased. We refer the reader to Hu et al. (2002) for the evolution of the
number of countries rated by Standard & Poor.
As mentioned above, country risk ratings encompass economic, financial and political aspects.
The statistical data of the economic and financial variables considered in this paper come from
the International Monetary Fund (World Economic Outlook database), from the World Bank
(World Development Indicators database) while those about the ratio of debt to gross domestic
product come from Moody’s publications. Values of political variables are provided by
Kaufmann et al. in two papers (1999a,b) that are joint products of the Macroeconomics and
Growth, Development Research Group and Governance, Regulation and Finance Institutes which
are affiliated with the World Bank. Before describing the relevance of the selected variables, we
discuss in Section 2.2 the selection method used.
The second criterion is the availability of complete and reliable statistics. We want to avoid
difficulties related to missing data that could reduce the statistical significance and the scope of
our analysis. For instance, according to recent information received from The World Bank6, their
research concentrates on developing economies and they have data on the debt of 137 countries to
whom they loan funds and who report their external debt to The World Bank. Since high income
countries do not receive World Bank funds, they do not report their debt numbers to The World
Bank. Such situations have significantly complicated the process of compiling complete debt
statistics. Hu et al. (2002) also report the problem of data availability.
The third criterion is the uniformity of data across countries. We have considered, for
example, incorporating the unemployment rate statistics disclosed by the World Bank. However,
the World Bank underlines that unemployment is analyzed and compiled according to definitions
which differ from country to country.
It is worth noting that in addition to the variables listed in Tables 13 and 14 (see Appendix),
Haque et al. (1996), Cantor and Packer (1996), Larrain et al. (1997), Monfort and Mulder (2000)
and Hu et al. (2002) use a dummy variable that represents the historical solvency of a country.
Haque et al. (1996) use the lagged rating at time (t-1) as an independent variable in their
regression model. Monfort and Mulder (2000) claim that membership in the OECD is likely to be
a significant indicator for country risk ratings. The same authors emphasize also the importance
of the location of countries, by adding to their set of independent variables two dummy variables
to characterize the country’s location in Asia or in Latin America. Hu et al. (2002) also use
regional dummy variables.
6
Anat Lewin, Private Communication, Development Data Group, The World Bank, June 06, 2001
7
Acronyms in parentheses following the name of variables are used in tables and appendices for referring to
variables.
8
Calculated on the basis of purchasing power parity in international dollars.
9
PAGE 10 RRR 08-2004
expenses, it must resort to inflationary money financing. High inflation rate results in a
substantial consumers’ purchasing power reduction and increases political discontent.
Trade balance (TB): trade balance is the balance of trade in goods expressed as a percentage of
GDP (purchasing power parity-PPP). This is the difference in value between a country's total
imports and exports (including information of oil and non oil exports, consumer goods, capital
goods) measured in current U.S. dollars divided by the value of GDP converted into international
dollars using purchasing power parity rates.
Exports’ growth rate (EGR): annual growth rate of exports of goods and services based on
constant local currency. Exports of goods and services represent the value of all goods and other
market services provided to the rest of the world. They include the value of merchandise, freight,
insurance, transport, travel, royalties, license fees, and other services, such as communication,
construction, financial, information, business, personal, and government services. They exclude
labor and property income as well as transfer payments. Countries having a high export growth
rate are expected to be more creditworthy. Indeed, exports are the primary source of foreign
currency inflows and therefore have a significant influence on the capacity of the country to
finance imports and service debt obligations.
International reserves (RES): this variable refers to gross international reserves, expressed in
terms of the number of months for which the existing reserves can cover the cost of imports of
goods and services. It gives an indication of the short-term capacity of an economy to meet its
imports obligations. The higher the value of RES, the lower the risk of default and the higher the
creditworthiness.
Fiscal balance (FB): fiscal balance is approximated by the ratio of central government financial
balance (surplus or deficit) to GDP. The central government’s balance represents the yearly fiscal
balance. Fiscal balances and debt stocks of governments are crucial indicators when analyzing
sovereign risk. The ability of governments to extract revenues from taxpayers and users of
services is a key factor that helps to determine whether governments will be able to make full and
timely payments of interest and principal on outstanding debt.
Debt to GDP (DGDP): here debt refers to the general government debt. The general government
debt as defined by the IMF (2001) includes “the consolidated budgets of the central,
state/regional, and local governments, along with the social security system and other extra-
budgetary funds engaged in noncommercial activities. Excluded are lending and refinancing and
the assets/liabilities of commercial state-owned or guaranteed enterprises, except for any net
financial transfers made as subsidies to these enterprises”. This balance, i.e., the difference
between total revenues and total expenditures, determines the net borrowing requirement of
general government, which can be met only by running down financial assets or borrowing net
new resources from the public and, thereby, adding to debt.
We have considered incorporating the unemployment rate and the ratio of the current account
balance to GDP. While the latter turned out to be redundant with trade as a percentage of GDP,
the former has been excluded from consideration due to the lack of consistency in its definition.
As noted by the World Bank, the treatment reserved to temporarily laid off workers, to those
looking for their first job, and the criteria referred to for being considered as unemployed, differ
significantly between countries.
For political variables, it is very difficult to find reliable and complete data. In our model, we
have considered the six variables provided by Kaufmann et al. (1999a). These six variables are:
political stability and violence, voice and accountability, government effectiveness, regulatory
burden, corruption , rule of law. These variables are viewed as capturing the fundamentals of the
RRR 08-2004 PAGE 11
governance concept defined as “the traditions and institutions by which authority in a country is
exercised” (Kaufmann et al.,1999a).
As emphasized by Kaufmann et al. (1999, a and b), political stability and voice and
accountability both refer to the process by which governments are elected, monitored and
replaced. Government effectiveness and regulatory burden reflect the capacity of the government
to adopt sound policies. Corruption and rule of law are proxies for the “respect of citizens and
institutions for the rules which govern their interactions”. In order to avoid or at least limit
redundancies in our model, we select only one variable for each dimension of governance. We
have selected:
Political stability (PS),
Government effectiveness (GE), and
Corruption (COR).
The higher the values of these variables, the less likely the country is to default9. The variables
are defined on a (-3.5, 3.5) interval and are based on estimations provided by polls of experts and
cross-country surveys.
The variables we have described so far have been considered previously in the literature and
are available in the form used in our study (as ratios or as growth rates). We have also decided to
construct a new variable (ER) and to add a variable (financial depth and efficiency) which, to the
best of our knowledge, has not been used before in country rating studies. Here are the
descriptions of these two variables:
Exchange rate (ER): is defined as the ratio of the current value of the exchange rate to the
moving average of the real effective exchange rate10 over five years (1994 to 1998). While the
exchange rate has been used in previous country rating studies, we consider the ratio introduced
here to be more significant, since it indicates the dynamics of changes in the exchange rate, by
specifying whether the trend is up (ER>1) or down (ER<1).
Financial depth and efficiency (FDE): is represented by the ratio of the domestic credit provided
by the banking sector to the GDP. Households accumulate claims on financial institutions that,
acting as intermediaries, pass funds to final users. Correlated to the development of the economy,
the indirect lending by savers to investors becomes more efficient and gradually increases assets
relative to the GDP. Viewed from this perspective, the ratio of domestic credit to the GDP
reflects the financial depth and efficiency of the country’s financial system. More specifically,
this variable is used to measure the growth of the banking system since it reflects the extent to
which savings are financial. To our knowledge, the financial depth and efficiency variable has not
been considered previously in the evaluation of country risk ratings.
• gross domestic product per capita, inflation rate, trade balance, international
involving nine economic/financial variables:
reserves, fiscal balance, exports growth rate, debt to GDP, financial depth and
9
The higher the value of the corruption variable, the less corrupted the considered country is perceived to be. This
variable can therefore be called “corruption quality”.
10
Real effective exchange rate is the nominal effective exchange rate (a measure of the value of a currency against a
weighted average of several foreign currencies) divided by a price deflator or index of costs.
11
PAGE 12 RRR 08-2004
efficiency, and exchange rate (we have used the values taken by these variables at
the end of 1998);
We have compiled the values of these twelve variables for the sixty-nine countries considered:
24 industrialized countries, 11 Eastern European countries, 8 Asian countries, 10 Middle Eastern
countries, 15 Latin American countries and South Africa. We use the Standard & Poor country
risk ratings for these countries at the end of December of 1998.
3 A Statistical Model
In order to derive a non-recursive model of Standard & Poor’s ratings, we shall fit the
Y = α + ∑ βi * X i + ε ,
regression equation:
M
(1.1)
i =1
where the dependent variable Y is the country risk rating given by Standard & Poor at the end of
December 1998 (or more precisely a numerical representation of Standard & Poor’s ratings), the
independent variables X i are the economic and political variables described in Section 2.1, and
ε is the error term. In view of the desired non-recursiveness of the model, the independent
variables do not include directly or indirectly ratings of previous years. Results given in this
section have been obtained using the SPSS statistical package.
The proposed model exhibits an excellent fit, with the coefficient of multiple determination R-
square being 91.2%, and the adjusted R-square 89.3%. The multiple correlation level between the
observed values (i.e. the Standard & Poor ratings) and the predicted ones (i.e. the ratings given by
the non-recursive regression model) is equal to 95.5%. The later are given in Appendix (Table 17,
Column 2).
Table 1 below details how the regression equation accounts for the variability in the response
variable, the last column giving the statistical level (1-p) at which the model is significant.
Table 2 presents the regular and the standardized regression coefficients (i.e., those
corresponding to the model fitted to standardized data). The last column in Table 2 indicates
whether the corresponding independent variable is statistically significant (at the confidence level
of 1-p).
Table 2: Regression results
Variables Unstandardized Standard error Standardized t-statistic p-value
coefficients coefficients
(Beta)
Intercept 8.769 .860 10.195 0.000
FDE 1.693E-02 0.007 0.148 2.513 0.015
RES 0.116 0.101 0.055 1.148 0.256
IR -2.831E-02 0.018 -0.080 -1.557 0.125
TB -1.192E-02 0.006 -0.094 -1.960 0.055
EGR -1,218E-02 0.031 -0.017 -0.396 0.694
GDPC 3.081E-04 0.000 0.499 5.294 0.000
ER -1.968E-02 0.011 -0.079 -1.768 0.083
FB 0.120 0.086 0.078 1.393 0.169
DGDP -1.610 0.795 -0.091 -2.026 0.047
PS 1.378 0.533 0,197 2.584 0.012
GEF 1.977 0.920 0.316 2.149 0.036
COR -0.605 0.842 -0.111 -0.718 0.476
At the 5% significance level, it appears that five independent variables are statistically
significant. These are: financial and depth efficiency (FDE), gross domestic product per capita
(GDPc), ratio debt to gross domestic product (DGDP), political stability (PS) and government
efficiency (GE).
The regression results described above indicate that the non-recursive regression model has an
excellent fit with the data. However, the excellence of the fit does not automatically guarantee the
predictive power of the model, if the model violates some of the critical assumptions of multiple
regression theory, as is the case with proposed model. Indeed,
• there is a strong correlation between some of the variables considered, e.g. between
the political variables, especially government efficiency and corruption, possibly
leading to difficulties related to multicollinearity, and the ill-conditioned nature of
the resulting matrix;
• the predictors are not normally distributed;
• if too many predictor variables are used relatively to the number of observations,
fitting multiple regression can lead to overfitting, and the estimates of the regression
line can be unstable and the results may not be reproducible; the number of
variables used is generally recommended to be no more than 5 to 10% of the
number of observations, which is clearly not the case of this study that involves 69
observations and 12 variables.
In view of these issues, it is surprising that the cross-validation results presented in the next
section provide a strong confirmation of the predictive power of the non-recursive regression
model presented above.
13
PAGE 14 RRR 08-2004
3.1.2 Cross-validation
To validate the predictive power of the non-recursive regression model, we use a resampling
technique known as cross-validation, and more specifically, a popular variant of it called k-
folding (e.g., Shao, 1993, Shao and Tu, 1995, Efron, 1982, Hurvich and Tsai, 1989, Hjorth 1994,
Breiman and Spector, 1992). In k-folding, observations are divided into k subsets of
approximately equal size. The regression model is trained k times, each time leaving out from
training one of the k subsets, and using the omitted subset to test the regression-predicted country
risk rating. In this paper, based on the relatively small size of the sample, we have selected k to be
10, and partitioned the sample into 10 groups of 6 or 7 countries each. The groups were selected
using stratified random sampling, i.e. assuring that each group contains about the same number of
investment-, speculative- and default-grade countries (see S&P’s classification, Table 15).
In Appendix (Table 17, Column 3) we present the in-the-sample predictions of the non-
recursive multiple regression model obtained in the preceding Section, and the out-of-the-sample
predictions obtained using the 10-fold cross-validation. The major results are the following:
• the correlation between the in-the sample and the out-of-the sample predictions is
• the correlation between the Standard and Poor ratings and the out-of-the sample
99.1%,
predictions is 95.6%.
The very high correlation levels demonstrate clearly that the impressive results of Section
3.1.1 are not due to chance or overfitting.
In this section, we shall identify the few countries for which the predictions of the non-
recursive regression model disagree with the Standard & Poor ratings. In order to accomplish
this, we shall construct confidence intervals for our predicted ratings11.
Let us introduce some notations. Let n and p refer to the number of observations and
predictors, respectively. The expression t (1 − α / 2, n − p ) refers to the Student test with
( n − p ) degrees of freedom, and with upper and lower tail areas of α / 2 . Let X j be the p-
dimensional vector of the values taken by the observation Y j on the p predictors, while X 'p be the
transposed of X j . Let the expression ( X ' X ) −1 refer to the variance-covariance matrix, i.e. the
inverse of the [ p × p ] -dimensional matrix ( X ' X ) . Denoting by MSE the mean square of errors in
^
the regression, the estimated variance s 2 [Y j ] of the predicted rating is:
{Y j − t (1 − α / 2, n − p ) * s[Y j ], Y j + t (1 − α / 2, n − p ) * s[Y j ]}
^ ^ ^ ^
(1.3)
11
All formulae given in this section as well as those in Section 3.2.3 are from Neter et al. (1996)
RRR 08-2004 PAGE 15
We say that there is a discrepancy between the Standard & Poor rating R SPj of a country j and
ours, if the Standard & Poor rating is not in the confidence interval, i.e.:
j ∉ {Y j − t (1 − α / 2, n − p ) * s[Y j ], Y j + t (1 − α / 2, n − p ) * s[Y j ]} for α = 0.1
^ ^ ^ ^
R SP (1.4)
Taking α equal 5%, this formula identifies four discrepancies. Three countries (Iceland,
Pakistan and Argentina) are rated higher by the non-recursive regression model than by Standard
& Poor, while Columbia is rated higher by Standard & Poor. It is remarkable that subsequently
the Standard & Poor ratings for two of these four countries (Columbia and Pakistan) have been
modified in the direction suggested by the regression model. More precisely, Columbia has been
downgraded by Standard & Poor twice, moving from BBB- in December 1998 to BB+ in
September 1999, and then to BB in March 2000. After being downgraded in January 1999 (SD),
Pakistan was upgraded to B- in December 1999. On the other side, Iceland’s rating has remained
unchanged, and Argentina’s rating has endured significant downgrade, which started however
only in November 2000.
In this section, we test the predictive power of the non-recursive regression model, from which
the three political variables are omitted. The R-square as well as the adjusted R-square of this
model are equal to 88.6 % and 86.9 % respectively. These values are lower than the
corresponding values for the original non-recursive regression model, indicating a loss in
predictive power resulting from the omission of the three political variables. The predicted ratings
are given in Appendix (Table 17, Column 4).
It appears that five of the independent variables are statistically significant at a 95% level.
These variables are financial and depth efficiency (FDE), gross domestic product per capita
(GDPc), debt to gross domestic product ratio (DGDP), exchange rate (ER) and fiscal balance
(FB). The correlation coefficient between the predicted ratings and those of Standard & Poor is
equal to 94.14 % and is lower than in the original model. Moreover, the inferior fit of this model
results in wider confidence intervals as compared to the original model.
15
PAGE 16 RRR 08-2004
The discrepancies between Standard & Poor’s predictions and ours involve four countries
(Russia, Pakistan, South Korea and Iceland), all being underrated by Standard & Poor. The
ratings of three of these countries (Russia, Pakistan, South Korea) have been modified since, in
the direction suggested by our model, while the rating of Iceland has remained unchanged. The
evolution of ratings for Pakistan has already been described in Section 3.1.3. South Korea has
both been upgraded three times, moving from BB+ in December 1998 to BBB+ in November
2001. Russia has first been downgraded to SD in January 1999, before being upgraded three
times and being rated B+ in December 2001.
In conclusion, the model which omits political variables appears to be somewhat less closely
related to the S&P model which it is supposed to reflect, but on the other end this apparent
weakening is not sufficiently clear to allow us to draw any definite conclusions. It should be
added here that the economic variables are easier to obtain than the political ones, which are
published less frequently.
In addition to analyzing the correlation level between Standard & Poor’s ratings and those of
the proposed non-recursive model, the latter has to be compared with the ratings of other
agencies, e.g. Moody’s and The Institutional Investor. We present below the results of these
comparisons, based on Moody’s and The Institutional Investor ratings issued at the end of
December 1998 and in March 1999 respectively. We shall start by presenting a brief description
of the rating systems of Moody’s and of The Institutional Investor.
Moody’s sovereign ratings are defined, as “a measure of the ability and willingness of the
country’s central bank to make available foreign currency to service debt, including that of
central government itself” (Moody’s, 1995). Similarly to Standard & Poor, Moody’s uses a
nominal rating scale (Table 18 in Appendix), which contains the same number of categories as
Standard & Poor’s ratings. A large proportion of countries receive the same rating from Moody’s
and Standard & Poor, and when they are different, the difference is usually not more than one
notch.
The Institutional Investor country risk ratings were first compiled in 1979, and are published
now regularly, in March and September of every year, for an increasing number of countries,
which reached 145 in 2000. The Institutional Investor ratings are numerical, ranging from 0 to
100, with 100 corresponding to the lowest chance of default. The Institutional Investor relies on
evaluations of the creditworthiness of the countries to be rated, provided by economists and
international banks, each respondent using their own criteria. Responses are aggregated by The
Institutional Investor, greater weights being given to responses from institutions with higher
worldwide exposure.
The correlation levels between the ratings given by the non-recursive multiple regression
model and those given by Standard and Poor, Moody’s and The Institutional Investor are reported
in Table 4.
RRR 08-2004 PAGE 17
17
PAGE 18 RRR 08-2004
{Y j ,n − t (1 − α / 2, n − p ) * s[ pred ], Y j ,n + t (1 − α / 2, n − p ) * s[ pred ]}
^ ^
(1.6)
SP
We say that there is a discrepancy between the Standard & Poor rating R j and the non-
recursive regression model, if :
j ∉ {Y j ,n − t (1 − α / 2, n − p ) * s[ pred ], Y j ,n + t (1 − α / 2, n − p ) * s[ pred ]} for α = 0.1
^ ^
R SP (1.7)
The results show that the only three discrepancies between our ratings and those of Standard &
Poor concern Argentina, Iceland and Russia, all of these countries being underrated by Standard
& Poor. The cases of Argentina and Iceland have already been discussed in Section 3.1.3. As far
as Russia is concerned, it was first downgraded to SD in January 1999, but upgraded afterwards
to B- in December 2000, and to B in June 2001, thus confirming our prediction.
In this section, we test the prediction power of our model on a set of countries which were not
used for constructing our model. We have obtained data for four such countries (Ecuador,
Guatemala, Jamaica and Papua New Guinea), the ratings of which by Standard & Poor started
after December 1998. The ratings of these countries using the 1998 non-recursive regression
model (described in Section 3) with 1998 and 1999 data are presented in Table 6.
Comparing the ratings predicted by the non-recursive regression model with those given by
Standard & Poor, it can be seen that the model has a significant predictive power, even when
applied to countries not used in constructing the model. Indeed, three of the four countries above
(Papua New Guinea, Jamaica and Guatemala) have their first Standard & Poor ratings within the
95% confidence intervals of the predicted ratings. The rating of Ecuador is even more interesting.
While the first Standard & Poor rating of that country (SD) is not in the 95% confidence interval
of the predicted value, that rating (given in July 2000) was revised after only one month (in
August 2000) to B, which falls within the 95% confidence interval of our prediction.
4 A Combinatorial Model
4.1 Pairwise country comparisons: Pseudo-observations
Let us associate to every country i ∈ I = {1,…,69} considered in this study, the 13-
dimensional vector Ci, whose first component is the country risk rating given by Standard and
Poor, while the remaining 12 components specify the values of the nine economic/financial and
of the three political variables.
19
PAGE 20 RRR 08-2004
pairs of countries. For this purpose, we shall construct for every pair of countries i, j ∈ I, a
In this study, instead of considering countries independently of each other, we shall consider
pseudo-observation Pij , which shall provide in a way specified below a comparative description
of the two countries.
The pseudo-observations are represented as 13-dimensional vectors. The first component is an
indicator which takes the value 1 if the country i in the pseudo-observation Pij has a higher rating
(i.e., lower risk) than the country j, takes the value –1 if the country j has a higher rating than the
country i, and takes the value 0 if both countries have the same rating. The other components k , k
= 2,…,13 of the pseudo-observation Pij[k] are obtained simply by taking the differences of the
Pij [k ] = Ci [k ] − Ci [k ], k = 2,...,13
corresponding components:
(1.8)
One of the advantages of this transformation is that it allows us to avoid the problems posed
by the fact that the original dataset contains only a small number (| I |) of observations. The
transformation (1.8) provides a substantially larger dataset, which contains | I |*(| I | -- 1) pseudo-
observations. It will be seen that the potential problems created by the non-independence of
pseudo-observations can be overcome by robust data analysis techniques.
The fundamental idea of this study is that a rating system can be essentially reconstructed from
the knowledge of the relative standings of all pairs of rated countries. In other words, all that
matters in a rating is the order relation between countries. Therefore, this study will focus on
inducing a model for the order relation between countries.
In order to illustrate the construction of pseudo-observations, let us consider as an example the
case of Japan and Canada.
Table 7 reports the values taken by the twelve economic/financial and political variables, as
well as the rating given by Standard & Poor to these countries at the end of December 1998.
Table 7: Examples of country observations
S&P
FDE RES IR TB EGR GDPc ER FB DGDP PS GE COR
RATING
CJapan AAA 138.44 5.168 .65 21.7471 -2.54 24314.2 0.839 -7.7 0.47 1.153 0.839 0.724
CCanada AA+ 94.69 1.01964 0.99 55.9177 8.79 24855.7 0.939 0.9 0.5 1.027 1.717 2.055
• a sufficiently high proportion of the negative (positive) observations violate at least one
satisfy the conditions imposed by the pattern, and
21
PAGE 22 RRR 08-2004
are satisfied by) each of the positive (negative) observations in the dataset. Furthermore, good
models tend to minimize the number of points in the dataset covered by both positive and
negative patterns in the model.
The way a LAD model can be used for classification is the following. An observation (whether
it is contained or not in the given dataset) which satisfies the conditions of some of the positive
(negative) patterns in the model, but which does not satisfy the conditions of any of the negative
(positive) patterns in the model, is classified as positive (negative). An observation satisfying
both positive and negative patterns in the model is classified with the help of a discriminant
which assigns specific weights to the patterns in the model (Boros et al., 2000). More precisely, if
p and q represent the number of positive and negative patterns in a model, and if h and k represent
∆ (θ ) = h / p - k / q, (1.9)
observation for which ∆ (θ ) = 0 is left unclassified, since the model either does not provide enough
and the corresponding classification is determined by the sign of this expression. Finally, an
evidence, or provides conflicting evidence; fortunately it has been seen in all the real-life
problems considered that the number of unclassified observations is extremely small.
we shall derive an LAD model, whose discriminant provides a numerical measure ∆(Pij) of the
countries having different S&P ratings which of the two is rated higher. From this information,
way that their pairwise differences provide the best approximation of the numerical measures ∆
countries, called “logical rating scores” (LRS); the logical rating scores are obtained in such a
obtained in the first step. These scores will be shown in Section 5 to have a very high correlation
with the S&P ratings.
The “observations” of the dataset used in the first step are those pseudo-observations Pij,
which correspond to countries i and j having different ratings. Each pseudo-observation Pij is
classified as positive or negative, according to the value of the indicator variable, i.e., depending
on whether i is rated higher than j or vice versa. Clearly, the training set is anti-symmetric.
∆ ( Pij ) for each pseudo-observation Pij (i ≠ j). The values ∆ ( Pij ) of the discriminant are called the
After having constructed the LAD model, we compute (according to (1.9)) the discriminant
relative preferences, and the [69 x 69]-dimensional anti-symmetric matrix, ∆, having them as
components will be called the relative preference matrix. While the LAD model was derived
matrix components are the values ∆ ( Pij ) (i ≠ j) taken by the discriminant for every pair of
using only those pseudo-observation Pij for which i and j were rated differently, the discriminant
To illustrate the concept of the relative preference matrix, let us consider the example of three
countries, Japan, Canada, and Belgium, and the six associated pseudo-observations. The derived
relative preferences obtained from the LAD model are shown in Table 9.
value of the relative preference ∆( Pi , j ) . A value equal or nearly equal to 0 would mean that the
creditworthy than country j , while the opposite conclusion could be drawn from a large negative
evidence for drawing conclusion about the relative creditworthiness of countries i and j is either
lacking or conflicting.
The difficulty of this naïve approach is that it overlooks the imprecision of (i.e., the noise
inherent in) data, and therefore of the relative preferences. The matrix above illustrates this
phenomenon, since the naïve interpretation would rate Japan above Canada, Canada above
Belgium, and at the same type Belgium above Japan, which contradicts the basic requirement of
transitivity of an order relation.
In the following section in order to overcome this difficulty, we shall relax the overly
constrained search for (possibly non-existent) country ratings whose pairwise orderings are in
precise agreement with the signs of relative preferences, to the more flexible search for logical
rating scores (LRS), having numerical values whose pairwise differences approximate well the
relative preferences.
4.3.2 From relative preferences to logical rating scores using regression analysis
It has been common practice in the research literature (see e.g., Ferri and Liu, 1999, Hu et al.,
2002, Kräussl, 2000, Monfort and Mulder, 2000, Sy, 2003) to interpret sovereign ratings as
cardinal values. Assuming that the sovereign ratings β can be interpreted as cardinal values, it is
natural to view the relative preferences ∆ as differences of the corresponding ratings:
∆( Pij ) = β i − β j , for all i, j ∈ I , i ≠ j (1.10)
∆(π ) = ∑ β k * xk (π ) +ε (π ) ,
∆’s can be found as a solution of the following multiple linear regression problem:
(1.11)
k∈I
π = {(i, j ) i, j ∈ I , i ≠ j}
where
(1.12)
23
PAGE 24 RRR 08-2004
1, for k = i
and
xk (i, j ) = −1, for k = j
0, otherwise
(1.13)
The logical rating scores β k estimated by the regression model are presented in column 4 of
Table 21 (in Appendix).
In order to evaluate the matrix ∆ of relative preferences obtained using LAD, we shall need a
comparable point of reference. A natural benchmark of this sort can be associated to any set of
numerical scores si representing sovereign ratings, by defining the canonical relative preferences
dij to be simply the differences dij = si – s j associated to every pair of countries i and j. In
this paper, we shall compare the LAD relative preferences ∆ij with the canonical relative
preferences dS&Pij , dMij , dIIij , dREGij , and dLRSij obtained respectively from the scores associated
with S&P’s ratings, Moody’s ratings, The Institutional Investor’s scores, the non-recursive
regression model scores, and the logical rating scores. The corresponding matrices of relative
preferences will be denoted dS&P , dM , dII , dREG , and dLRS respectively. The correlation levels
are shown in Table 10, where the symmetric correlation matrix is filled out completely, in order
to make comparisons easier.
The high levels of correlation show that the relative preferences obtained using LAD are in a
surprisingly good agreement both with the ratings of S&P and those of the other agencies, as well
as with the non-recursive regression scores. A comparison of the logical rating scores with the
LAD relative preferences shows the clear superiority of the former, in view of their higher degree
of agreement with the other canonical relative preferences.
We shall analyze now the correlation levels between the logical rating scores and the scores
associated with the ratings of S&P, Moody and The Institutional Investor, as well as those
provided by the non-recursive regression model (columns 2, 10 and 11, respectively, in Table 21
in the Appendix). It can be shown (Appendix) that these correlation levels are exactly identical to
those between the corresponding canonical relative preference matrices, which are presented in
Table 10.
Several valuable conclusions can be derived from the correlation levels reported in Table 10.
First, the high levels of correlation between the scores show that both the logical rating scores and
those provided by the non-recursive regression model are very good approximations of the S&P
ratings, as well as of those provided by other rating agencies.
Second, it is remarkable that, in spite of the entirely different nature of the LRS and the non-
recursive regression techniques, their results have an extremely high correlation (98.29%).
Moreover, the correlation between one of these two scores and any individual agency rating, is
almost identical to the correlation between the other score and the rating of that agency.
This similarity is striking due to several essential distinctions between LRS and the non-
recursive regression technique. The first distinction is that the regression technique assumes that
the differences between consecutive ratings are the same, while the LRS is not based on any
assumption about the magnitude of the differences. The second – and perhaps most striking
distinction – concerns the mathematical techniques used for deriving the two scores: the non-
recursive regression score is obtained through a classical statistical technique, while the LRS is
derived through combinatorial-logical concepts and techniques.
The third distinction concerns the mathematical functions describing the relationship between
the independent variables and the scores: while the non-recursive regression scores depend
linearly on the variables, the LRS depends on them in a complex nonlinear fashion, through the
intermediary constructs of logical patterns.
The most important conclusion derived from this discussion is the fact that the almost identical
results, provided by two techniques of essentially different natures, strongly reinforce each other.
It has been seen in the previous section that the LRS and the S&P ratings are in close
agreement. However, since the logical rating scores and the S&P ratings are not expressed on the
same scale, the comparison of the two scores of an individual country presents a challenge. In
order to overcome this difficulty, we shall apply a linear transformation to the LRS, which brings
them to the same scale as the S&P ratings. This is accomplished by determining the coefficients a
and c for the transformation a*βi + c of the LRS βi in such a way that the mean square difference
between the transformed LRS and the S&P ratings is minimized. Obviously, these coefficients
can be determined by simple linear regression. As a result of this transformation, the LRS become
25
PAGE 26 RRR 08-2004
directly comparable with the S&P ratings. Clearly, the consistency of (i.e., the correlation
coefficient between) the LRS and S&P ratings is not affected by this transformation.
Using formula (1.6), we then compute the confidence intervals for the transformed LRS of
each country in the dataset. In 1998, five countries have a S&P rating that does nor fall within the
confidence interval of the transformed LRS. Columbia appears to be too favorably rated by S&P,
while Hong-Kong, Malaysia, Pakistan and Russia appear to be rated too harshly by S&P. As
explained in Sections 3.1.3 and 3.1.4, the evolution of the S&P ratings for Columbia, Pakistan
and Russia is in agreement with the 1998 LRS of these countries, and underlines the prediction
capability of the LRS model. It is remarkable that the evolution of the S&P ratings of Malaysia
and Hong-Kong is also in agreement with their 1998 LRS. Indeed, both Malaysia and Hong-
Kong have been upgraded shortly thereafter, the former moving from BBB- to BBB in November
1999, and the latter from A to A+ in February 2001.
In this section, we shall construct logical rating scores based on the 1999 data, in order to
evaluate the robustness and consistency of the LRS model. This goal will be accomplished by
comparing the LRS obtained in this way with Standard and Poor’s 1999 ratings. The logical
ratings scores for the 1999 data are given in Appendix (Table 21, columns 8).
It has been seen in Section 4.4.2 that the scores provided by the two models (the non-recursive
regression and the LRS models) built on the 1998 Standard & Poor ratings when applied to the
1998 data are strikingly similar. In this section, we compare the results of the same two methods
applied to the 1999 data.
RRR 08-2004 PAGE 27
The first conclusion resulting from the high levels of pairwise correlations between the S&P
1999 ratings, the relative preferences given by the LAD discriminant, and the canonical relative
preferences corresponding to non-recursive regression scores and LRS, is that both the LRS and
the non-recursive regression model have a very strong temporal stability. The second conclusion
is that the logical rating scores are superior to the relative preferences given by the LAD
discriminant.
The major conclusion resulting from this table is that the logical rating and the non-recursive
regression scores remain strikingly similar, and each of them provides a strong reinforcement of
the other.
4.5.2 Discrepancies
Using formula (1.6), we compute the confidence intervals for the transformed LRSe of each
country in the dataset. Applying the 1998 LRS model to the 1999 data, we see that only two
countries (Russia and Hong-Kong) have S&P ratings that are outside the confidence intervals of
the corresponding transformed LRS. These two countries appear to be rated too harshly by S&P.
As explained in Sections 3.1.4 and 4.4.3, the evolution of the S&P ratings for Hong-Kong and
Russia is in agreement with the 1999 LRS of these countries.
The availability of the LAD discriminant, which does not involve in any way the previous
years’ S&P ratings, makes it possible to rate previously non-rated countries in the following way.
After calculating first the attribute values of all the pseudo-observations involving the new
countries to be evaluated, the relative preferences are to be calculated for these pseudo-
observations, and the resulting columns and rows have to be added to the matrix of relative
preferences. The new LRS for all the countries (new and old) should then be determined by
running the multiple linear regression model (1.12).
In order to evaluate the capability of LRS to correctly predict S&P ratings, we compare the
LRS predicted as described above, with the S&P ratings when they first become available. The
direct comparison between these ratings is carried out using the linear transformation described in
Section 4.4.3.
Using formula (1.6), we compute the confidence intervals for the transformed LRS of four
countries never rated by S&P by December 1998. It appears that our predictions for three of them
(Guatemala, Jamaica and Papua New Guinea) correspond perfectly to the first time (subsequent)
S&P ratings. The comparison between the LRS and the first S&P rating (SD) given in July 2000
for the fourth country (Ecuador) shows that S&P rated it much too harshly, since one month later
S&P significantly raised its rating to B-, thus fully justifying the LRS prediction.
27
PAGE 28 RRR 08-2004
The correlation levels presented in Table 12 indicate that the matrices of relative preferences
∆JK and dJKLRS obtained through the jackknife procedure are highly correlated with (i) the
canonical relative preferences corresponding to the S&P ratings dS&P , and (ii) the in-the-sample
relative preferences ∆, as well as the corresponding logical rating scores dLRS . This shows that ∆
and dLRS are not affected by overfitting.
5 Concluding Remarks
The central objective of this paper was to develop transparent, consistent, self-contained, and
stable country risk rating systems, closely approximating the country risk ratings provided by a
major rating agency (Standard & Poor). We proposed in this paper two models that achieve the
stated objectives, the first one utilizing the classical econometric technique of multiple linear
RRR 08-2004 PAGE 29
regression, and the second one using the combinatorial-logical technique of Logical Analysis of
Data.
The proposed models are highly accurate, having a 95.5% correlation level with the actual
S&P ratings and almost equally high correlations with Moody’s and The Institutional Investor.
The models avoid overfitting, as demonstrated by the 99.1% correlation between in- and out-of-
sample rating predictions, calculated by k-fold cross-validation. The proposed models are
transparent since they make the role of the economic-financial and political variables explicit.
The proposed models are distinguished from the rating models in the existing literature by
their self-contained nature, i.e., by their non-reliance on any information derived from lagged
ratings. Therefore, the high level of correlation between predicted and actual ratings cannot be
attributed to the reliance on lagged ratings, and is a reflection of the relevance and predictive
power of the independent variables included in these models. The significant advantage of the
non-recursive nature of the proposed models is their applicability to not-yet-rated countries.
The consistency of the proposed models is illustrated by the fact that the few discrepancies
between the S&P ratings and those of the proposed models were resolved by subsequent changes
in S&P’s ratings. The stability of the constructed non-recursive models is shown in two ways: by
predicting 1999 ratings using the non-recursive models derived from the 1998 dataset applied to
the 1999 data, and – most importantly -- by successfully predicting the ratings of several
previously non-rated countries, i.e., in full agreement with subsequent ratings of those countries
by S&P.
The confidence in the results and in the validity of both models is strongly reinforced by the
fact that the traditional linear regression model and the qualitatively different combinatorial-
logical model produce almost identical results.
The significance of the results of this paper is further confirmed in a forthcoming study
(Hammer et al., 2004), in which the LAD discriminant constructed in this paper is used for
deriving a partial order representing the creditworthiness of countries, and then extending this
partial order to a new rating system. The comparison of the results of the present and the
forthcoming studies shows that – in spite of yet another qualitatively different data analysis
approach -- the predicted ratings are strongly correlated, at the surprising level of 98-99%.
29
PAGE 30 RRR 08-2004
References
Altman E.I. and Saunders A. 1997. Credit Risk Measurement: Developments over the Last 20
Years. Journal of Banking and Finance 21 (11-12), 1721-1742.
Aylward L. and Thorne R. 1998. Countries’ Repayments Performance Vis-à-Vis the IMF”, IMF Staff
Papers 45 (4), 595-619.
Bhatia A.V. 2002. Sovereign Credit Risk Ratings: An Evaluation. IMF Working Pape WP/
03/170.
Boros E., Hammer P.L., Ibaraki T. and Kogan A. 1997. Logical Analysis of Numerical Data,
Mathematical Programming 79, 163-190.
Bouchet M.H., Clark E. and Groslambert B. 2003. Country Risk Assessment: A Guide to
Global Investment Strategy. John Wiley & Sons Ltd. Chichester, England.
Bourke P. and Shanmugam B. 1990. An Introduction to Bank Lending. Addison-Wesley
Business Series.
Breiman L. and Spector P. 1992. Submodel Selection and Evaluation in Regression: The X-Random
Case. International Statistical Review 60, 291-319.
Brewer T.L. and Rivoli P. 1990. Politics and Perceived Country Creditworthiness in
International Banking. Journal of Money, Credit and Banking 22, 357-369.
Brewer T.L. and Rivoli P. 1997. Political Instability and Country Risk. Global Finance
Journal 8 (2), 309-321.
Cantor R. and Packer F. 1996. Determinants and Impact of Sovereign Credit Ratings. FRBNY
Economic Policy Review, 37-53.
Citron J.T. and Neckelburg G. 1987. Country Risk and Political Instability. Journal of
Development Economics 25, 385-395.
Claessens S. and Embrechts G. 2002. Basel II, Sovereign Ratings and Transfer Risk: External
versus Internal Ratings. Presentation at the Basel II: An Economic Assessment, Bank for
International Settlements, Basel, Switzerland.
Cook W.D. and Hebner K.J. 1993. A Multicriteria Approach to Country Risk Evaluation:
With an Example Employing Japanese Data. International Review of Economics and Finance 2
(4), 327-348.
Coplin W.D. and O’Leary M.K. 19983. Political Risk Yearbook, The PRS Group, New York.
Cosset J.C. and Roy J. 1991. The Determinants of Country Risk Ratings. Journal of
International Business Studies 22, 135-142.
Cosset J.C., Siskos Y. and Zopounidis C. 1992. Evaluating Country Risk: A Decision Support
Approach. Global Finance Journal 3, 79-95.
Crama Y., Hammer P.L. and Ibaraki T. 1988. Cause-Effect Relationships and Partially
Defined Boolean Functions. Annals of Operations Research 16, 299-326.
Cruces J.J. 2003. Statistical Properties of Sovereign Credit ratings. Working Paper. University
of San Andres, Buenos Aires.
Doumpos M. and Zopounidis C. 2001. Assessing Financial Risks Using a Multicriteria
Sorting Procedure: The Case of Country Risk Assessment. Omega 29, 97-109.
RRR 08-2004 PAGE 31
Easton S.T. and Rockerbie D.W. 1999. What’s in a Default ? Lending to LDCs in the Face of
Default Risk. Journal of Development Economics 58 (2), 319-332.
Eaton J., Gersovitz M. and Stiglitz J.E. 1986, The Pure Theory of Country Risk. European
Economic Review 30, 481-513.
Efron B. 1982. The Jackknife, the Bootstrap and Other Resampling Plans. Philadelphia, SIAM.
Erb C.B., Harvey C.R. and Viskanta T.E. 1996a. Expected Returns and Volatility in 135
Countries. Journal of Portfolio Management, 46-58.
Erb C.B., Harvey C.R. and Viskanta T.E. 1996b. Political Risk, Economic Risk and Financial
Risk. Financial Analysts Journal, 29-46.
Estrella A. 2000. Credit Ratings and Complement Sources of Credit Quality Information. Basel
Committee on Banking Supervision Working Paper 3.
Feder G. and Uy V.U. 1985. The Determinants of International Creditworthiness and their
Policy Implications. Policy Modeling 7 (1), 133-156.
Ferri G., Liu L-G. and Stiglitz J. 1999. The Procyclical Role of Rating Agencies: Evidence
from the East Asian Crisis. Economic Notes 3, 335-355.
Hammer P.L. 1986. Partially Defined Boolean Functions and Cause-Effect Relationships.
International Conference on Multi-Attribute Decision Making Via OR-Based Expert Sytems.
University of Passau, Passau, Germany.
Hammer P.L., Kogan A. and Lejeune M.A. 2004. Modeling Country Risk Ratings Using
Partial Orders. RUTCOR Technical Report 1-2004, Rutgers University, New Brunswick, New
Jersey.
Haque N.U., Kumar M.S., Mark N. and Mathieson D. 1996. The Economic Content of
Indicators of Developing Country Creditworthiness. International Monetary Fund Working
Paper 43 (4), 688-724.
Haque N.U., Kumar M.S., Mark N. and Mathieson D. 1998. The Relative Importance of
Political and Economic Variables in Creditworthiness Ratings. International Monetary Fund
Working Paper 46, 1-13.
Haque N.U., Mark N. and Mathieson D. 1997. Rating the Raters of Country Creditworthiness.
Finance & Development 34, 10-13.
Hjorth J.S.U. 1994. Computer Intensive Statistical Methods Validation, Model Selection, and
Bootstrap. London, Chapman & Hall.
Hu Y.-T., Kiesel R. and Perraudin W. 2002. The Estimation of Transition Matrices for
Sovereign Credit Ratings. Journal of Banking and Finance 26 (7), 1383-1406.
Hurvich C.M. and Tsai C.-L. 1989. Regression and Time Series Model Selection in Small Samples.
Biometrika 76, 297-307.
Jüttner J.D. and McCarthy J. 2000. Modeling a Rating Crisis. Macquarie University,
Unpublished, Sidney, Australia, http://www.econ.mq.edu.au/staff/djjuttner/SOVEIG2.pdf.
Kaminsky G. and Schmukler S.L. 2002. Emerging Market Instability: Do Sovereign Ratings Affect
Country Risk and Stock Returns? World Bank Economic Review 16, 171-195.
Kaufmann D., Kraay A. and Zoido-Lobaton P. 1999a. Aggregating Governance Indicators.
World Bank Policy Research Department Working Paper 2195,
http://www.worldbank.org/wbi/governance/pdf/agg_ind.pdf.
Kaufmann D., Kraay A. and Zoido-Lobaton P. 1999b. Governance Matters. World Bank
31
PAGE 32 RRR 08-2004
Larrain G., Reisen H. and von Maltzan J. 1997. Emerging Market Risk and Sovereign Credit
Ratings. OECD Development Center, 1-30.
Lee S.H. 1993. Relative Importance of Political Instability and Economic Variables on
Perceived Country Creditworthiness. Journal of International Business Studies 24, 801-812.
Levich R., Reinhart C., and Majnoni G. 2002. Ratings, Rating Agencies and the Global
financial System. Kluwer Academic Press. New York.
Manasse P., Roubini N. and Schimmelpfennig A. 2003. Predicting Sovereign Debt Crises.
IMF Working Pape WP/ 03/221.
Mathieson D. and Schinasi G. 1999. International Capital Markets: Developments, Prospects,
and Key Policy Issues. IMF Series: World Economic and Financial Surveys.
Mauro P. 1993. Essays on Country Risk, Asset Markets and Economic Growth. Harvard
University Dissertation.
Mauro P. 1998. Corruption and the Composition of Government Expenditure. Journal of
Public Economics 69 (2), 263-279.
Monfort B. and Mulder C. 2000. Using Credit Ratings for Capital Requirements on Lending
to Emerging Market Economies: Possible Impact of a New Basel Accord. IMF Working Paper
WP/00/69.
Moody’s. 2001. Moody’s Country Credit Statistical Handbook, First Edition, New York..
Mulder C. and Perelli R. 2001. Foreign Currency Credit Ratings for Emerging Market
Economies. IMF Working Paper WP/01/191.
Mulekar M.S. and Mishra S.N. 2000, Confidence Interval Estimation of Overlap: Equal
Means Case, Computational Statistics & Data Analysis 34, 121-137.
Bilson C. M., Brailsford T.J. and Hooper V.J. 2001. Selecting Macroeconomic Variables as
Explanatory Factors of Emerging Stock Market Returns. Pacific-Basin Finance Journal 9 (4),
401-426 .
Neter J., Kutner M.H., Nachtsheim C.J. and Wasserman W. 1996. Applied Linear Statistical
Models. Fourth Edition. Irwin/McGraw-Hill.
Nickell P., Perraudin W. and Varotto S. 2000. Stability of Rating Transitions. Journal of
Banking and Finance 24, 203-227.
Nordal K.B. 2001. Country Risk, Country Risk Indices and Valuation of FDI: A Real Options
Approach. Emerging Markets Review 2, 197-217.
RRR 08-2004 PAGE 33
Oral M., Kettani O., Cosset J.-C. and Daouas M. 1992. An Estimation Model for Country
Risk Rating. International Journal of Forecasting 8 (4), 583-593.
Quenouille, M. 1949. Approximate Tests of Correlation in Time Series. Journal of the Royal
Statistical Society, Series B 11, 18-84.
Quenouille, M. 1956. Notes on Bias in Estimation. Biometrika 43, 353-360.
Reinhart C.M. 2002. Default, Currency Crises, and Sovereign Credit Ratings. World Bank
Economic Review 16, 151-170.
Reisen H. and von Maltzan J. 1998. Sovereign Credit Ratings, Emerging Market Risk and
Financial Market Volatility. Intereconomics 33 (2), 73-82.
Reisen H. and Von Maltzan J. 1999. Boom and Bust and Sovereign Ratings. International
Finance 2 (2), 273-293.
Saini K.G. and Bates P.S. 1984. A Survey of the Quantitative Approaches to Country Risk Analysis.
Journal of Banking and Finance 8, 341-356.
Shao J. 1993. Linear Model Selection by Cross-Validation. Journal of the American
Statistical Association 88, 486-494.
Shao J. and Tu D. 1995. The Jackknife and Bootstrap. New York, Springer-Verlag.
Simpson J. 2000. The World Trade Organisation and Basel Accord Membership and
Compliance in Relation to International Banking Risk. Curtin University Working Paper 35.
Somerville R.A. and Taffler R.J. 1995. Banker Judgment versus Formal Forecasting Models :
The Case of Country Risk Assessment. Journal of Banking and Finance 19 (2), 281-297.
Standard & Poor. 1999. Sovereign Ratings Display Stability over Two Decades. Credit Week.
Standard & Poor. 2000. Ratings Performance. Standard & Poor. New York.
Standard & Poor. 2001. Sovereign Ratings History since 1975. New York.
Sy A.N.R. 2003. Rating the Rating Agencies: Anticipating Currency Crises or Debt Crises?
IMF Working Paper WP/ 03/122.
Tang J.C.S. and Espinal C.G. 1989. A Model to Assess Country Risk. OMEGA: International
Journal of Management Sciences 17 (4), 363-367.
World Bank. 2000a. Global Development Finance 2000. Washington D.C.
World Bank. 2000b. World Development Indicators 2000. Washington D.C.
33
PAGE 34 RRR 08-2004
Appendix
Table 13:Economic variables and literature
Variable Literature
Consumer price index Larrain et al. (1997), Hu et al. (2002)
Credit claims on central government
Monfort and Mulder (2000)
growth rate
Haque et al. (1996,1998), Larrain et al. (1997), Brewer and Rivoli
Current account balance / GDP (1990), Doumpos et al. (2001), Cosset et al.(1992), Cook and
Hebner (1993), Tang and Espinal (1990)
Larrain et al. (1997), Feder and Uy (1985), Aylward and Thorne
(1998), Dailami and Leipziger (1997), Cosset and Roy (1991),
Debt12 / exports Cosset et al. (1992), Cantor and Packer (1996), Lee (1993),
Doumpos and Zopounidis (2001), Monfort and Mulder (2000), Hu
et al. (2002), Mulder and Perelli (2002)
Haque et al. (1996, 1998), Feder and Uy (1985), Lee (1993),
Debt13 / GDP Brewer and Rivoli (1990), Aylward and Thorne (1998), Doumpos
and Zopounidis (2001), Cook and Hebner (1993), Hu et al. (2002)
Debt / reserves Monfort and Mulder (2000), Manasse et al. (2003)
Dependence on oil exportation Feder and Uy (1985)
Domestic investment / GDP Larrain et al. (1997), Monfort and Mulder (2000)
Exports / GDP Aylward and Thorne (1998)
Exports concentration Feder and Uy (1985)
Haque et al. (1996,1998), Feder and Uy (1985), Doumpos and
Exports growth rate Zopounidis (2001), Cosset et al.(1992), Monfort and Mulder
(2000)
Exports variability Cosset et al.(1992)
Exports vulnerability to external
Feder and Uy (1985)
shocks
Larrain et al. (1997}, Brewer and Rivoli (1990), Monfort and
External debt / GDP
Mulder (2000), Manasse et al. (2003)
Larrain et al. (1997), Cantor and Packer (1996), Lee (1993),
Fiscal balance14
Monfort and Mulder (2000) , Cook and Hebner (1993)
Foreign investment policy Cook and Hebner (1993)
Haque et al. (1996,1998), Larrain et al. (1997), Feder and Uy
(1985), Cantor and Packer (1996), Doumpos and Zopounidis
GDP15 growth rate
(2001), Monfort and Mulder (2000), Cook and Hebner (1993), Hu
et al.(2002)
Dailami and Leipziger (1997), Erb et al. (1997), Feder and Uy
GDP per capita (1985), Cosset et al. (1991, 1992), Larrain et al. (1997), Monfort
and Mulder (2000), Tang and Espinal (1990)
GDP per capita growth rate Lee (1993), Haque et al. (1996), Aylward and Thorne (1998)
Easton and Rockerbie (1999), Cosset et al. (1991, 1992),
Gross investment / GDP
Doumpos and Zopounidis (2001)
12
Same as for footnote 6.
13
The word “debt” can encompass foreign, total, debt service or external debt, depending on authors.
14
Central government spending / GDP, domestic public debt / GDP and are used as a proxy for this variable.
15
Note that authors use GDP as well as GNP.
RRR 08-2004 PAGE 35
Variable Literature
Anti-governmental demonstrations Haque et al. (1998)
Armed conflicts (or riots) Haque et al. (1998), Brewer and Rivoli (1990), Cook and Hebner
(1993)
Assassination Haque et al. (1998)
Corruption Mauro (1993)
Coups Haque et al. (1998)
General strikes Haque et al. (1998)
Guerilla warfare Haque et al. (1998)
Influence of the middle class Mauro (1993), Cook and Hebner (1993)
Legal system Mauro (1993)
Major government crises Haque et al. (1998)
Political change Mauro (1993), Brewer and Rivoli (1990)
Political legitimacy Brewer and Rivoli (1990)
Political stability Brewer and Rivoli (1990, 1997), Feder and Uy (1985), Citron and
Neckelburg (1987), Mauro (1993), Lee (1993), Cosset et al.(1992),
16
Represented by a dummy variable.
35
PAGE 36 RRR 08-2004
Level Description
An obligor rated AAA has extremely strong capacity to meet its financial
AAA
commitments. AA is the highest issuer credit rating assigned by S&P.
An obligor rated AA has very strong capacity to meet its financial
INVESTMENT
AA
commitments. It differs from the highest rated obligors only in small degree.
RATING
An obligor rated A has strong capacity to meet its financial commitments but is
A somewhat more susceptible to the adverse effects of changes in circumstances
and economic conditions than obligors in higher-rated categories.
An obligor rated BBB has adequate capacity to meet its financial commitments.
However, adverse economic conditions or changing circumstances are more
BBB
likely to lead to a weakened capacity of the obligor to meet its financial
commitments.
An obligor rated BB is less vulnerable in the near term than other lower-rated
SPECULATIVE
commitments.
RATING
Table 16: Standard & Poor’s country ratings (end of December, 1998)
We have converted the Standard & Poor rating scale (columns 1 and 4) into a numerical scale
(columns 2 and 5). Such a conversion is not specific to us. Bouchet et al. (2003), Estrella (2000),
Ferri et al.(2001), Kräussl (2000), Monfort and Mulder (2000), Mulder and Perelli (2001), Sy
[2003] proceed similarly. Moreover, Bloomberg, a major provider of financial data services,
developed a standard cardinal scale for comparing Moody’s, S&P and Fitch-BCA ratings
(Kaminsky and Schmukler, 2002). A higher numerical value denotes a higher probability of
default. The numerical scale is referred to in this paper as Standard & Poor’s preorder.
37
PAGE 38 RRR 08-2004
39
PAGE 40 RRR 08-2004
SPECULATIVE RATING
Aaa
obligations
Aa1 High quality Ba2
Aa2 Ba3 Ongoing uncertainty
Aa3 B1 High risk obligations
INVESTMENT
A2 B3
Caa Current vulnerability to
A3
default or in default
DEFAULT
RATING
Adequate payment Ca In bankruptcy or default.
Baa1
capacity
Baa2 D
Baa3
Table 19: Standard & Poor’s country risk ratings: average one-year transition rates (1975-1999)
41
PAGE 42 RRR 08-2004
c = ∑∑ cij / n 2 = 0 , d = ∑∑ d ij / n 2 = 0,
n n n n
(1.17)
i =1 j =1 i =1 j =1
∑ (ai − a )(bi − b )
1 n
ρ ( a , b) =
n i =1
∑ i ∑ (bi − b )2
(1.18)
−
1 n 2 1 n
( a a )
n i =1 n i =1
∑∑ (c − c )( d ij − d ) ∑∑ (a − a j )(bi − b j )
n n n n
1 1
ρ (C , D ) =
i =1 j =1 i =1 j =1
ij i
n2 n2
∑∑ (c ∑∑ (d ∑∑ (a ∑∑ (b − b )
= (1.19)
− c) − d) − aj)
n n n n n n n n
1 2 1 2 1 2 1 2
i =1 j =1 i =1 j =1 i =1 j =1 i =1 j =1
ij ij i i j
n2 n2 n2 n2
∑∑ (a − a j )2 can be rewritten as
n n
The expression
i =1 j =1
i
∑ ∑ [(a − a j )2 + ( a j − ai )2 ] = 2∑ ∑ ( ai − a j )2
n −1 n n −1 n
(1.20)
i =1 j = i +1 i =1 j = i +1
i
1
PAGE 2 RRR 08-2004
Using (1.20) and (1.21), the correlation between the matrices C and D (1.10) can be rewritten
as :
∑ ∑ (a
n −1
− a j )(bi − b j )
n
1
ρ (C , D ) =
i =1 j = i +1
i
n2
∑ ∑ (a ∑ ∑ (b − b )
n −1 n −1
(1.22)
− aj)
n n
1 2 1 2
i =1 j = i +1 i =1 j = i +1
i i j
n2 n2
∑ i n∑ − − = ∑∑ i j i − − = ∑∑(ai − a j )(ai − a )
1 n 1 n 1 n n 1 n n
( a a )( a a ) ( (a a ))( a a )
n i =1 j =1 n2 i =1 j =1 n2 i =1 j =1
j i
2 ∑∑
[(ai − a j )(ai − a ) +(a j − ai )(a j − a )] = 2 ∑ ∑ (ai − a j )(ai − a −a j + a ) = 2 ∑ ∑ (ai − a j )2 (1.23),
1 n−1 n 1 n−1 n 1 n−1 n
=
n i =1 j=i+1 n i=1 j=i+1 n i=1 j=i+1
∑ ∑ [(a ∑ ∑ (a
n −1 n −1
= − a j )( bi − b ) + ( a j − ai )( b j − b )] = − a j )( bi − b − b j + b ) (1.24)
n n
1 1
i =1 j = i +1 i =1 j = i +1
i i
n2 n2
∑ ∑ (a
n −1
= − a j )(bi − b j )
n
1
i =1 j = i +1
i
n2
Using (1.23) and (1.24), the correlation between the vectors a and b (1.19) can be rewritten as
∑ ∑ (a
n −1
− a j )(bi − b j )
n
1
ρ ( a , b) =
i =1 j = i +1
i
n2
∑ ∑ (a ∑ ∑ (b − b )
n −1 n −1
, (1.25)
− aj )
n n
1 2 1 2
i =1 j =i +1 i =1 j = i +1
i i j
n2 n2